Skip to content

worker: instructions lead with what memory is for; claim tool names feedback as what accepts - #87

Merged
vu1n merged 8 commits into
mainfrom
claude/project-thread-ghkdsb
Nov 7, 2026
Merged

vu1n merged 8 commits into
mainfrom
claude/project-thread-ghkdsb

Conversation

@vu1n

@vu1n vu1n commented Nov 7, 2026

Copy link
Copy Markdown
Owner

Requested by Vu · project thread

The MCP initialize instructions open with what memory is for (what the next agent would otherwise pay to rediscover; not status or what the code says) and keep each step to a line or two in plain case, instead of five capitalised procedures. The claim step asks for evidence (#86) and says an untested guess is not a lesson.

Also fixes drift: the claim tool description said a candidate is "accepted once another session claims the same subject", which contradicts the echo guard; it now names a used rating via feedback. The kypp README template and AGENTS.md block mention evidence.

Why: Anthropic's agent memory post found scripted write rules cut memory gains for capable models while a plain statement of purpose helped. kypp has no eval for instruction wording, so this is a hypothesis to watch in the usefulness report, not a measured win.

Contracts touched: initialize instructions text; claim tool description.

Verification: npm run typecheck; npm test 388/388 (instruction assertions in feedback.test.ts updated; read-only callers still never see claim); featuremap.py check clean.


Generated by Claude Code

@vu1n vu1n self-assigned this Nov 7, 2026
@vu1n
vu1n merged commit 65a3933 into main Nov 7, 2026
4 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant